基于注意力机制和可分离卷积的双目立体匹配算法  

Stereo matching based on attention mechanism and separable convolution

作　　者：郑秋梅[1] 温阳 王风华[1] ZHENG Qiu-mei;WEN Yang;WANG Feng-hua(Department of Computer and Communication Engineering,China University of Petroleum,Qingdao 266580,China)

机构地区：[1]中国石油大学(华东)计算机科学与技术学院,山东青岛266580

出　　处：《微电子学与计算机》2021年第5期42-47,共6页Microelectronics & Computer

基　　金：国家自然科学基金(61305008);中央高校基本科研业务费专项资助(19CX02030A)。

摘　　要：针对当前基于卷积神经网络的双目立体匹配算法需要较高的特征提取能力且网络的参数量过多的问题,提出一种基于注意力机制的立体匹配网络,在特征提取阶段采用改进后的通道注意力机制根据通道内所含的信息进行特征加权,同时采用改进的空间金字塔结构实现多尺度特征提取,以提高网络的特征提取能力;设计3D注意力模块和3D可分离卷积进行视差计算,相比于标准卷积不仅可以降低网络的计算参数同时增加了通道维数可以保证匹配精度.最后,在Scene Flow数据集、KITTI 2012数据集和KITTI 2015数据集上进行评估,实验表明,本文的网络模型在保证匹配精度的同时有效减少网络的计算参数.To improve the feature extraction capability of current stereo matching network and reduce the parameter,the multi-scale context attention network for stereo matching network is proposed.In the feature extraction stage,the improvedchannel-wise attention mechanism is used to weight the features based on the information contained in the channel,and the improved spatial pyramid structure is used to achieve multi-scale feature extraction to improve the network’s feature extraction ability.A three-dimensional attention module and a three-dimensional separable convolution for disparity calculation.Compared with the standard convolution,the computational parameters of the network can be reduced and the channel dimension can be increased to ensure the matching accuracy.Experiments on Scene Flow datasets,KITTI 2012 datasets and KITTI 2015 datasets show that the network model in this paper can effectively reduce the calculation parameters of the network while ensuring matching accuracy.

关 键 词：双目视觉 立体匹配 注意力机制 可分离卷积 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]